Scalable surrogate deconvolution for identification of partially-observable systems and brain modeling
Matthew F Singh1,2,3, Anxu Wang1,2, Todd S Braver2
1Department of Electrical and Systems Engineering, Washington University in St. Louis, St. Louis, MO, United States of America.
Journal of Neural Engineering
|June 27, 2020
Summary
Surrogate deconvolution reconstructs biological system activity from indirect measurements. This scalable method accurately models brain networks and physiological signals, outperforming current standards.
Area of Science:
- Biophysics
- Computational Neuroscience
- Systems Biology
Background:
- Direct in-vivo measurement of all biophysical system state-variables is often infeasible.
- Reconstructing hidden system activity from indirect measurements is a key challenge in biological modeling and signal processing.
- Indirect measurements frequently arise from linear time-invariant dynamical interactions, representable as a convolution of hidden states with an unknown kernel.
Purpose of the Study:
- To introduce surrogate deconvolution, a novel approach for identifying coupled biophysical systems and parameterizing models.
- To reframe nonlinear, partially observable identification problems in neuroscience and biology as analytical objectives.
- To provide a method compatible with various optimization procedures for system identification.
Main Methods:
- Development and application of the surrogate deconvolution technique.
- Comparison of surrogate deconvolution with joint Kalman Filters (Unscented and Extended) for partially observable system estimation.
- Application to simulations of Local Field Potential (LFP) and blood oxygen level dependent (BOLD) signals.
- Empirical stability assessment of Hemodynamic Response Function (HRF) kernel estimates for Mesoscale Individualized NeuroDynamic (MINDy) models.
Main Results:
- Surrogate deconvolution is highly scalable and computationally efficient.
- The technique performs competitively with established methods like joint Kalman Filters.
- Demonstrated benefits for model identification using LFP and BOLD signal simulations.
- Empirically stable HRF kernel estimates were obtained for individual human brain MINDy models, showing reliable individual variation and stereotyped spatial distribution.
Conclusions:
- Surrogate deconvolution offers a powerful tool for parameterizing coupled biophysical systems.
- The method enhances brain-modeling by enabling rapid, simultaneous fitting of large-scale network models and physiological processes.
- This approach promises to advance neuroscientific measurement analysis, particularly for BOLD fMRI hemodynamics.
